Delegate a Discovery
Some code needs more than one attempt to improve, such as a slow inner loop or a request path whose latency you want down. You can hand it to a Discovery Run from your coding agent and keep working while it runs.
What happens
You or your agent set the Objective and point Artemis at the Project, its Runner and its Script. The Discovery Run then works inside Artemis, so your agent and your terminal stay free. It generates many candidate Versions with several models, measures each one on your Runner against the Baseline, and builds on the promising ones. Several Discovery Runs can go at the same time.
Later, ask your agent how it is going. It can list the Versions, compare the best one with the Baseline Metric by Metric, show the diff, and steer the run by adding budget or giving it a new direction.
Before you merge
The best Version still needs your review. Read the diff, check that the improvement is larger than the noise, and make sure the benchmark reflects real use. Many candidates fail or show no gain, which is normal. You decide what gets merged.